What data for this category looks like
Daily documents for this category include packaging process cards, bonding and molding parameters, wafer incoming material drawings, test program versions, yield reports, and failure analysis reports. Data is sourced from wafer test maps, MES systems, bonding and molding equipment, ATE testers, X-ray/SAM inspection equipment, and customer specification requirements. Data is updated per wafer batch, per individual chip test result, and daily equipment operating parameters. Each document includes fields such as equipment parameter number, batch number, yield value, and failure type, with units matching the corresponding measurement standards for the document type.
What constraints do these characteristics impose on model access and configuration
The multi-source heterogeneous business data sources of this category require model access to support structured extraction and association of cross-system data, and configuration must account for format compatibility across different data sources. The high-frequency update rhythm requires context configuration to adapt to real-time or near-real-time data synchronization requirements, avoiding batch alignment errors caused by expired context. The bidding scenario requires strict retention of original document parameters and batch information. Model output must not arbitrarily modify character formats or field units. It is also necessary to address resource occupancy issues under concurrent requests to avoid excessive latency. In addition, local model deployment requirements are common in Linux environments, and access rules adapted to local APIs must be configured.
The configuration below is formulated based on the business characteristics of semiconductor packaging and testing, and is compatible with FastGPT v4.17.0.
How to set the configuration
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | The business data of this category is multi-source and updated by batch/daily, so sufficient historical batch and test parameter context must be retained to avoid loss of key alignment information |
temperature | 0.1–0.3 | The bidding scenario requires strict matching of original document formats and parameters, so divergent content generation should be avoided to maintain output consistency |
quoteMaxToken | 1500–2000 token | Long document content such as wafer incoming material drawings and test program versions must be fully referenced to avoid truncation of key parameters or batch identifiers |
AIPROXY_API_ENDPOINT | http://localhost:11434/v1 | Adapted to large model calls for local ollama deployment in Linux environments, reducing public network transmission latency |
maxTokens | 2500–3500 token | The output length of a single bid document and quotation must cover the complete packaging plan and quotation details to avoid forced truncation of content |
| Workflow/API tool selection | auto | Automatically trigger document reference tools to align wafer maps and physical batch information, no manual configuration of call rules required |
Three common mistakes
- Phenomenon: A request error is returned after configuring a local reranking model. Cause: The API address and authentication parameters of the reranking model are not correctly matched, or the network access permission for local models is not enabled in FastGPT.
- Phenomenon: When concurrent requests increase, model answer latency increases significantly, and background logs show frequent queries and writes to MongoDB. Cause: The reasonable value of
maxContextis not limited, causing each request to load excessive historical data and triggering high-frequency read and write operations on MongoDB. - Phenomenon: When calling a large model, the output content has extra spaces or letters are changed to uppercase. Cause:
temperatureis not set to a low range, and strict format constraints are not configured, causing the model to arbitrarily adjust the character format of the original document.
How to confirm successful configuration
- Initiate a single test request, and check whether the output content fully references the configured business document fields, with no arbitrarily modified character formats or parameter errors.
- Simulate concurrent requests, observe background resource usage, and confirm that latency does not increase irregularly as the number of concurrency increases.
- Check the model API call logs to confirm that the request address and authentication parameters match the configured items, with no invalid network errors.
- Verify the connectivity of the local model deployment, confirm that the model service in the Linux environment can be normally accessed by FastGPT, with no connection timeout issues.
The values provided here are common starting points and should be measured against the reader's own samples.